Write once, rewrite everywhere: A Unified Framework for Factorized Machine Learning

This thesis describes TRINITY, a framework to optimize linear algebra algorithms operat- ing over relational data in GraalVM. The framework implements a host-language-agnostic version of the optimizations introduced by the Morpheus project, meaning that a single implementation of the Morpheus rewrite rules can be used to optimize linear algebra algorithms written in arbitrary GraalVM languages. We evaluate its performance when hosted within FastR and GraalPython, GraalVM’s R and Python implementations respectively. In doing so, we also show that TRINITY can optimize across languages, meaning that it can execute and optimize an algorithm written in one language, such as Python, while using data originating from another language, such as R